Request: Lightning LoRA optimized for Qwen Image Edit 2509 FP8 non-scaled model
#33
by
Mkeni
- opened
When editing any image using the Qwen Image Edit 2509 FP8 (non-scaled variant) model together with any Lightning LoRA, a grid-like pattern appears in the output. Using the scaled FP8 version provided in the repository resolves this issue, but inference becomes approximately twice as slow (20s/it) compared to the non-scaled variant (around 10s/it). What’s surprising is that it’s even slower than the full BF16 model with offloading (around 12 s/it). My hardware is an RTX 3060 and 64 GB of DDR4 RAM.
Would it be possible to release a dedicated set of Lightning LoRA weights optimized specifically for the non-scaled FP8 Image Edit 2509 base model, similar to what was done for the Qwen Image (non-edit) model?